Perception, prediction, and penetration
Zoe Drayson
(University of Stirling)
Abstract
The traditional view in philosophy and psychology maintains that perception and
cognition are distinct, and that perception can influence cognition but not vice versa.
This view is challenged in a new paper by Andy Clark, in which he suggests that
perception and cognition are ‘intertwined’, ‘continuous’, and ‘unified’, and that
perception is ‘theory-laden’. These claims about the nature of perception and
cognition, according to Clark, are the result of the hierarchical and predictive nature
of the human brain. In this paper, I show that this model of the brain is entirely
consistent with the traditional view of perception and cognition. In particular, the
debate over the cognitive penetrability of perception is left untouched.
Perception and cognition
Perception and cognition, according to the traditional view, play distinct roles in our
mental lives. Perception gathers information about the immediate environment;
cognition combines this with stored information (e.g. beliefs, memories) to draw
inferences, and with motivational states (e.g. desires) to guide action. The relation
between the two is traditionally understood to be one-way: what we perceive
influences what we believe, but our perception of the world is not (directly)
influenced by our beliefs. Another way to put this latter claim is to say that
perception is cognitively impenetrable – the contents of our perceptual experience
are not penetrated by our cognitive (doxastic, epistemic) states.
The standard case for the cognitive impenetrability of perception involves optical
illusions like the Muller-Lyer illusion, in which you are shown two lines that appear
(due to contextual effects) to be differing lengths. Even once you have measured the
two lines and convinced yourself that the two lines are the same length, you still
perceive them to be different lengths: your beliefs do not influence your perceptual
experience.
If the traditional view is false, then perception is cognitively penetrable. This would
have substantive philosophical implications regarding the epistemic role played by
perceptual experience. For example, rational scientific progress seems to require
that we can use observation to adjudicate between between competing scientific
theories, which in turn requires that observation itself be theory-neutral. But if
perception is cognitively penetrable, then one’s theoretical commitments can
influence what one perceives, and observation is ‘theory-laden’.
1
(Notice that as standardly understood, cognitive penetration is the penetration of
perceptual experience by conceptual thought. Top-down effects on ‘early vision’
processes from within the visual system would not count as cognitive penetration
for standard purposes: see Macpherson 2012 for discussion.)
Hierarchical predictive models
The relation between perception and cognition, therefore, impacts on key
philosophical issues. The traditional view of perception and cognition has recently
been challenged by Andy Clark (in press), who claims that a new model of
perception represents “a genuine departure” from our previous thinking about
perception and cognition. Clark suggests that this is a new model not only of
perception, but “perhaps ultimately of the relation between perception and
cognition itself”, in which perception and cognition are “profoundly unified and, in
important respects, continuous” and in which “[b]elieving and perceiving emerge as
deeply intertwined”.
The model to which Clark refers is the hierchical prediction model of brain function
(hereafter HPM). The key idea behind HPM is that the brain is ultimately a
predictive engine, composed of a hierarchy of systems operating according to
Bayesian rules. Higher-level systems try to predict the inputs to lower-level systems,
and any prediction errors are used by the higher-level system to refine its
predictions. The brain attempts to ‘explain away’ the input errors by generating
predictions that match the input. The higher-level guesses act as Bayesian priors for
the lower-level processing. As a model of perception, this means that the brain
predicts what the next sensory stimuli will be. Sensory input is not used as raw data
from which to build up a model of the environment: rather it is used to correct our
existing model of the environment.
Clark claims that HPM leads to a view of perception as (i) inferential, (ii) knowledgedriven, and (iii) the result of top-down processes. I’ll take each of these claims in
turn in at attempt to establish the extent to which they are true, and the extent to
which they make us rethink the relation between perception and cognition.
In what sense is perception inferential in HPM?
Clark claims that HPM follows the work of Helmholtz and others “in depicting
perception as a process of probabilistic, knowledge-driven inference”. The
‘inferences’ in question are Bayesian, and so, claims Clark, "[t]he process of
perception is thus inseparable from rational (broadly Bayesian) processes of belief
fixation". Claims about the ‘inferential’ nature of perception, however, are nothing
new. While some of these claims refer to genuinely inferential processes, most are
using the term metaphorically. As Hatfield points out, many people use ‘inferential’
as synonmous with ‘information-processing’: "They are not inferential theories of
perception, but theories of information transmission in perception" (Hatfield 2002).
2
So in what sense are Bayesian perceptual systems inferential? First, notice that
Bayesian models do not attribute perceptual inferences to the person: "There is no
good sense in which the thinker herself, as opposed to her perceptual system,
executes perceptual inferences” (Rescorla forthcoming). On the Bayesian view, the
transitions between perceptual states are “statistical inferences” in the sense that
"[t]ransitions among perceptual states approximately conform to norms of Bayesian
decision theory" (Rescorla forthcoming).
And indeed, Bayesian frameworks are a way of taking talk of unconscious inferences
are converting it into “mathematically rigorous, quantitatively precise psychological
models" (Rescorla forthcoming). Clark himself seems to be committed to such a
view of inference talk: the ‘free energy minimization’ framework is supposed to
discharge talk of prediction, inference, etc. leaving it harmlessly metaphorical.
So the fact that perceptual processing is Bayesian and 'inferential' in form doesn't
tell us anything new about the relation between perception and cognition. In
particular, perception is no more 'continuous with cognition' on this view than in
any other computational theory of perception.
To what extent is perception knowledge-driven in HPM?
Clark thinks that HPM gives us a picture of perception as involving "knowledgedriven inference." In what sense are the Bayesian statistical inferences in HPM
‘knowledge-driven’? Knowledge comes into the picture due to the predictive
commitments of the model, and the phenomenon of ‘explaining away’: higher-level
models attempt to cancel out prediction errors from sensory signals, to match the
predicted data to the actual data. Clark emphasises that "[t]hese predictions reflect
what the system already knows about the world”, and that "[t]o perceive the world
just is to use what you know to explain away the sensory signal” (emphasis mine).
First, notice that ‘knows’ here does not refer to agent-accessible propositional
attitude states involving belief and justification. 'Knowledge' here refers to the
assumptions of the visual system, where even the term ‘assumption’ is a
metaphorical gloss. Clark’s example of an assumption is the visual system’s
‘knowledge’ that only one object can exist in the same place at the same time. This is
not the agent’s belief about how the world is, but rather a prior assignment of
probabilities that acts as an input to Bayesian processing.
What should we make, then, of Clark’s claim that perception is “theory-laden in a
very profound way” on HPM? Clark suggests that perception is theory-laden in the
sense that what we see is determined by our best hypothesis. But this need not be
understood as 'our' (i.e. the agent's) best hypothesis, rather than the visual system's
best hypothesis. Compare the 1980s debate between Churchland and Fodor, in
which Churchland claimed that perception was theory-laden on the grounds that the
perceptual system involves certain assumptions. Fodor countered that these
3
assumptions leave “perception neutral with respect to almost all theoretical
disputes” and “couldn't ground any general argument for the unreliability of
observation” (Fodor 1988:189). I suggest that the same can be said for Clark’s
theory ladeness. What we perceive is partly determined by assumptions in the
visual system (either hard wired or acquired) such as that light generally comes
from above – but this has been a standard feature of theories of perception for a
long time due to the poverty of the stimulus. A genuine case of theory ladeness
would be the Muller-Lyer lines appearing the same length after we have learned
their measurements, but there is no indication that HPM gives us that. I suggest that
Clark’s ‘assumptions’ are like Churchland’s assumptions in that they “have nothing
to do with beliefs, theories, or other doxastic commitments that we may have [...]
[They] leave the epistemic normativity of perception untouched" (Stokes
forthcoming).
In what sense is perception top-down in HPM?
Even if we dismiss the sense in which perception is inferential or knowledge-driven
in HPM as being any different from other models of perception, there is a sense in
which the HPM clearly departs from other models in its emphasis on top-down
processing. The predictive hierarchies mean that higher hypotheses determine what
we see, with lower processing playing less of a role than usually supposed. But the
distinction between ‘top down’ and ‘bottom up’ effects is problematic, if ‘top down’
is taken to include implicit assumptions or expectations:
"There is a problem, however, with the psychological distinction. Top-down
effects are defined as influences of an individual’s expectations and stored
knowledge (Eysenck, 1998, p. 152). On the face of it, that includes the
expectations that are implicit in the operation of a psychological process – in
its dispositions to transition from one representation to another." (Shea
forthcoming)
The philosophically interesting cases of top-down effects are restricted to
information that is explicitly represented in the system. If we redefine the notion of
‘top down’ to retain the philosophically interesting distinction, as Shea does, then
the mere fact of hierarchical processing in a downwards direction is insufficient to
raise problems about cognitive penetration. The top-down influence has to come
from explicit representations – but Bayesian perceptual models don’t require
explicit representation:
"There is no evidence that the perceptual system explicitly represents
Bayes's Rule or expected utility maximization. The perceptual system simply
proceeds in rough accord with Bayesian norms." (Rescorla forthcoming)
Conclusion
Clark (forthcoming) claims that hierarchical predictive models of brain function
have implications for the relation between perception and cognition. Specifically, he
claims that perception is inferential, knowledge-driven, and top-down. I have
4
examined the extent to which perception fulfils these criteria, according to HPM, and
concluded that the traditional view of perception and cognition remains unaffected.
A further interesting question relates to the implications of HDM for a causal theory
of perception and for questions about realism, which I will discuss in a future paper.
Bibliography
Clark, Andy (in press) Whatever next? Predictive brains, situated agents, and the
future of cognitive science. Behav. Brain Sci.
Hatfield, Gary (2002) Perception as Unconscious Inference. In D. Heyer (ed.),
Perception and the Physical World: Psychological and Philosophical Issues in
Perception. John Wiley and Sons Ltd.
Macpherson, Fiona (2012) Cognitive Penetration of Colour Experience: Rethinking
the Issue in Light of an Indirect Mechanism. Philosophy and Phenomenological
Research 84 (1):24-62.
Rescorla, Michael (forthcoming) Bayesian Perceptual Psychology. In Mohan Matthen
(ed.) The Oxford Handbook of the Philosophy of Perception. Oxford.
Shea, Nicholas (forthcoming) Distinguishing Top-Down From Bottom-Up Effects. In
S. Biggs, M. Matthen & D. Stokes (eds.), Perception and Its Modalities. Oxford
University Press.
Stokes, Dustin (forthcoming) Cognitive penetrability of perception. Philosophy
Compass.
5